app.py

LANGSMITH_TRACING=True
LANGSMITH_ENDPOINT="https://api.smith.langchain.com"
LANGSMITH_API_KEY="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"
LANGSMITH_PROJECT="pr-complicated-disagreement-8"

import gradio as gr
from google import genai
import json
from langsmith import traceable

from IPython.display import display, HTML, Markdown


def show_json(obj):
  print(json.dumps(obj.model_dump(exclude_none=True), indent=2))
  return json.dumps(obj.model_dump(exclude_none=True), indent=2)

def show_parts(r):
  parts = r.candidates[0].content.parts
  if parts is None:
    finish_reason = r.candidates[0].finish_reason
    print(f'{finish_reason=}')
    return
  for part in r.candidates[0].content.parts:
    if part.text:
      display(Markdown(part.text))
      output = part.text
    elif part.executable_code:
      display(Markdown(f'```python\n{part.executable_code.code}\n```'))
      output = part.executable_code
    else:
      show_json(part)

  grounding_metadata = r.candidates[0].grounding_metadata
  if grounding_metadata and grounding_metadata.search_entry_point:
    display(HTML(grounding_metadata.search_entry_point.rendered_content))
  return output



client = genai.Client(api_key="AIzaSyD6voSAiSUim17kB90skpdisMMyFXZPxMo")
MODEL_ID = "gemini-2.0-flash-thinking-exp"

"""
For more information on `huggingface_hub` Inference API support, please check the docs: https://huggingface.co/docs/huggingface_hub/v0.22.2/en/guides/inference
"""
#client = InferenceClient("HuggingFaceH4/zephyr-7b-beta")
'''
import google.generativeai as genai2
genai2.configure(api_key=os.environ["GOOGLE_API_KEY"])
model_gen = genai2.GenerativeModel(model_name="gemini-2.0-flash-thinking-exp",
      generation_config=generation_config,
      system_instruction=system_instruction,
      safety_settings=safety_settings)
'''
def model_response(text):
   response = client.models.generate_content(
    model=MODEL_ID,
    contents = text
)
   return response.text
search_tool = {'google_search': {}}
code_tool = {'code_execution':{}}
tools = [
    {'google_search': {}},
    {'code_execution': {}}
]
soccer_chat = client.chats.create(model="gemini-2.0-flash-exp", config={'tools': [search_tool]})
coder_chat = client.chats.create(model="gemini-2.0-flash-exp", config={'tools': [code_tool]})

def agentville(problem):
  memory = {}
  output = ""
  final_response = ""
  plan = model_response(f'''You are a thinker. Think long and hard about the problem: {problem} and come up with the steps required to solve the problem. You are supposed to come up with 3 steps to build a solution.''')
  print ("Plan", plan)
  output += plan
  yield output, final_response
  for i in range(1,6):
    print ("Step:", i)
    output += f"Step {i} "
    yield output, final_response
    step = model_response(f'''Extract the {i}th step from the given plan: {plan}.  Figure out which of the below agents are required to solve it. 
    Once you figure out the agent, just output the name of the agent.
Do not output anything else. The agents at your disposal are:
Researcher: Has access to Google search tool, can search for resources and answer the questions
Coder: Expert programmer. Can solve any problem at disposal. Output the name of the agent and nothing else. 

Your response should be in the following format:
  "Step": <the complete description of the step to be executed>
  "agent_name": "Researcher/Coder"
  ''')
    print ("Current step", step)
    output += step
    yield output, final_response
    if 'Coder' in step:
      print ("Agent is coder")
      r = coder_chat.send_message(f'''Complete the step:{step}''')
      execution_step = show_parts(r)
      output += execution_step
      yield output, final_response
    else:
      print ("Agent is Researcher")
      r = soccer_chat.send_message(f'''Complete the step: {step}''')
      execution_step = show_parts(r)
      output += execution_step
      yield output, final_response
    memory[i] = execution_step  
  final_response = model_response(f'''Given the problem statement:{problem} and the progress made by agents: {memory}, come up with the final answer. Do not explain what 
  the agents have done. Focus on getting the final answer.''')
  print ("Final response", final_response)
  output += final_response
  yield output, final_response
  return output, final_response

import gradio as gr
iface = gr.Interface(
    fn=agentville,
    #inputs=["text", "text", "text"],
    inputs = gr.Textbox(label="Problem"),
    outputs= [gr.Textbox(label="Agents processing"),
            gr.Markdown(label="Final response")],
    title="Agentville: Power Up Your World ⚡️🌍 with Autonomous Agents 🤖💡 (Research Preview)",
    description="Watch multiple agents collaborate seamlessly to tackle complex challenges",
    theme = gr.themes.Ocean(),
    article = "All the conversations would be logged for quality improvement",
    examples = ["A white paper on the evolution of Gen AI","Explain how Quantum computing can disrupt our life","Is SAAS really dead in the wake of recent Gen AI developments?"]
)

# Launch the Gradio app
if __name__ == "__main__":
    iface.queue(max_size=20).launch(share=True,debug=True)